✦ 银河系系主任 ✦

Out of interest / Internet Cafe Attendant

[Boris Cherny @bcherny]

We talk a lot about how important it is to set up self-verification loops. Especially in the age of powerful models that can run for long periods of time, self-verification is a key ingredient that enables the model to run for much longer, delivering a result that is closer to what you intended, so you can do more without having to constantly check in on Claude as it works.

@delba_oliveira gives a great breakdown of what that looks like and why it matters

[Hwang @hwwaanng]

给大家一些 /goal 的用例，防止 token 用不完

1/ 产品说明文档

基于当前 commit 的产品功能，运行真实 App / 模拟器 / 使用 CDP 进行页面检查和截图，产出一份面向普通用户的产品说明文档。

文档需要用简洁清晰的语言解释每个主要页面、按钮、操作路径和点击后的行为，让用户能理解功能怎么用、为什么有用。

该文档应作为后续功能更新的持续记录基础，并在完成后沉淀为可复用的 Skill，方便之后按同一流程继续生成或更新产品说明。

验收标准

- 成功启动虚拟机或本地运行环境，并打开当前项目 App。
- 使用 CDP 或浏览器调试能力获取关键页面截图。
- 覆盖当前 commit 中已有的主要功能、页面和交互。
- 逐项说明按钮、入口、状态变化和点击后的结果。
- 文案面向普通用户，避免技术实现细节，表达简洁清楚。
- 文档能持续追加后续功能更新，而不是一次性说明。

最后总结并沉淀为一个可复用 Skill，包含运行、截图、梳理、撰写和更新文档的标准流程。

[✦ 银河系系主任 ✦ @fly_7_7]

@rwayne 咋不说只有三天病假🤡年假还是按时间逐步释放

[✦ 银河系系主任 ✦ @fly_7_7]

随地大小拉💩之前清理过一波又来了

[✦ 银河系系主任 ✦ @fly_7_7]

相似的光景三年前也经历过，但这次叠加组织的变化，更加扑朔迷离

[✦ 银河系系主任 ✦ @fly_7_7]

@suisuizone @st7evechou 移动端能搞不

[✦ 银河系系主任 ✦ @fly_7_7]

@JamesssWang @st7evechou 不好用，规则一多，经常反复连接断开连接断开，无语了

[✦ 银河系系主任 ✦ @fly_7_7]

@zzgo_sh 刚看到说修复了而且重置了https://t.co/0oJReNsQbC

[Tibo @thsottiaux]

Some of you noticed limits drained faster in Codex, we root caused it to an optimization that we rolled back that had an impact on cache hit rates when compacting across long running sessions.

We fixed this and have now reset usage limits for all accounts. Enjoy the weekend.

[✦ 银河系系主任 ✦ @fly_7_7]

gpt-5.5 high 相比 xhigh 就是个傻子

[✦ 银河系系主任 ✦ @fly_7_7]

试了下，居然调用了一个之前没见过的 skill：
/anthropic-skills:consolidate-memory

[Frad @FradSer]

一个 Claude Code 小技巧：

深度整理记忆。

Agent 的 LTM 是最好用的设计，但是记忆本身有问题会影响 agnet 的任务完成质量，也会浪费 token。

[✦ 银河系系主任 ✦ @fly_7_7]

A社很鸡贼(･_･;，
上一次重置的日期就在周四，所以对于很多爱用户来说是重置与否没啥区别

[歸藏(guizang.ai) @op7418]

Claude 重置了五小时和这周的额度限制，周末可以爽了

[九原客 @9hills]

经过好几周的摸索和各种尝试，pi agent 的配置基本稳定，之前发了一个 list，没有解释不太友好，这次发个全量带注释的。

避雷：如果需要开箱即用的 Coding Agent，请不要用 Pi，Claude Code/Codex 更适合你。

[Muharrem Şenyıl @senyil]

https://t.co/H8VuWhdzJ3 just got a big update!

- Waay faster
- A new deck every single day
- Explore by slide type (intro, pricing, team…)
- No membership, no paywall

Free, forever.

[歸藏(guizang.ai) @op7418]

https://t.co/lHQkxeTcRm

[岚叔 @LufzzLiz]

教程和开源代码已分享
https://t.co/zWXnQrGdBI

[Matt Pocock @mattpocockuk]

A flow I just tried and LOVED:

1. /grill-with-docs, talking about a new bit of UI
2. Asks me a question I can't answer unless I prototype
3. /prototype
4. Iterate on the prototype, burning tokens freely until we get a good spot
5. /rewind to the question, and select 'summarize' (Claude Code feature), saying 'summarize what we learned from prototyping'
6. Continue the grilling session, retaining the prototype

Smoooooooth

[✦ 银河系系主任 ✦ @fly_7_7]

这个问题很早就有讨论，有兴趣可以读读
https://t.co/H4Ovat12ls
https://t.co/aC01fho8X3
https://t.co/BTDWMali4K

[SkyWT @skywt2003]

为什么现在很多 UI 按钮开关都不做鼠标 hover pointer（变成手指）效果了？

shadcn 默认就是这样，codex 默认也是这样，甚至加了个设置项自定义这个。

[✦ 银河系系主任 ✦ @fly_7_7]

合理怀疑为了流量才发表这种类似串子的言论😅

[花叔 @AlchainHust]

我每天和Claude Code一起工作10小时，说一个「反共识」观点，其实Anthropic一直在骗你～

Skills中的Skill.md和Claude.md文档其实都是：
越长效果越好！
越长效果越好！
越长效果越好！

是的，Claude Code创始人Boris @bcherny 以及Anthropic官方经常出来告诫开发者说这两个文档应该写得短，尽量100行、200行以内。

但事实完全不是这么回事：
1、我在实践中发现如果你通过渐进式披露的策略把部分规则放在子文档，或者ref文档里，你的任务在比较复杂或者执行步骤长了之后根本不会被读取；
2、你可以去了解下不管是Claude Code还是Claude Design，抑或是Claude网页版的System Prompt有多长？如果写得简短是最佳实践，他们自己就不会这么干了...
3、他们之所以希望开发者这么干，是因为当一个人有几十上百个skill或者无数claude.md文档的时候，大家都短一些对skill间的协作有利，以及对他们System prompt发挥更大作用有利。

但这根本不重要！！你希望的是你自己的skill或者写给agent的规则能发挥到最大重要！！

所以，只要你用的模型足够好，指令遵循能力足够强，知道自己在做什么。是完全可以把文档写长的，一点问题都没有。

[Matt Pocock @mattpocockuk]

Tons of folks are piling in here saying that AFK agents are a myth.

I have been using them to ship these GitHub repos:

mattpocock/evalite
mattpocock/sandcastle
mattpocock/software-factory (might be public by the time you see this)

Here are a few steps to making this work, and some reality checks.

Definitions

Let's split this into the day shift and the night shift. Day shift is planning/review/QA, night shift is AFK implementation.

Day Shift (part 1)

1. Use /grill-me to align with the AI
2. Use /to-prd and /to-issues to create a PRD (the destination) and implementation steps as separate tickets, which can be grabbed in parallel (the journey)
3. The PRD is a ticket, but it's not an actionable step. You just put the user stories there

This is pure requirements gathering shit, same as it ever was.

Night Shift

1. I run a planner agent which looks at all the tickets and sees what can be worked on now, and what's blocked
2. The planner agent then kicks off multiple agents (sandboxed using Sandcastle, my OSS tool) to implement the code
3. I then have an automated reviewer agent look at the commits produced - one agent per implementation. This checks alignment to the original PRD, as well as code quality
4. These commits end up on branches that get PR'd to main
5. The planner agent runs again until all work has been completed

The review is a crucial step - it's saved me MANY times. I am planning to massively increase the amount of review I do, hopefully with multiple agents.

But guess what - AFK agents sometimes produce bad code. This can happen because of:

a. The original plan was bad because the best solution was something different
b. The original plan was bad because it didn't take into account all the unknown unknowns, and the AI had to make some decisions during the coding session which were bad
c. The plan was good, but the AI just shat the bed (twice, once in the review stage, once during implementation)
d. Your codebase is bad and the feedback loops don't tell the agent if it did a good job or not

So... QA:

Day Shift (part 2)

1. QA all of the branches created
2. Create follow-up issues, potentially editing the original PRD to adjust the destination

This will usually take a long time, often as long as planning. But then you kick off the night shift again.

Once QA is all done, you review the important bits of code manually, usually in PR's. There isn't anything better than the PR UI right now, so that's what we're stuck with.

Wake-up Calls

1. If you let the AI run all night unbounded by planning, it's going to produce shit code
2. Mostly, my loops finish before I go to bed, it's just the night shift catching up to the day shift
3. The only reason I do AFK at all is because it allows me to automate review and totally not give a shit about latency
4. I always run night and day shift in parallel. I can't plan that far ahead (skill issue, probably). I need working code to base my plans from, so I'm aggressively QA-ing stuff that lands

[✦ 银河系系主任 ✦ @fly_7_7]

官方现在出来修还是因为 5.5发布了，而且这不是第一回了，垃圾 A 社

[图拉鼎 @tualatrix]

官方承认降智了，我就一个问题：你们的模型这么厉害，怎么连降智这种问题都发现不了？还要一个多月来解决？